# Distillation(蒸馏)
## 内容
- [YOLOv3模型蒸馏](#YOLOv3模型蒸馏)
- [FGD模型蒸馏](#FGD模型蒸馏)
- [CWD模型蒸馏](#CWD模型蒸馏)
- [LD模型蒸馏](#LD模型蒸馏)
- [PPYOLOE模型蒸馏](#PPYOLOE模型蒸馏)
- [引用](#引用)
## YOLOv3模型蒸馏
以YOLOv3-MobileNetV1为例,使用YOLOv3-ResNet34作为蒸馏训练的teacher网络, 对YOLOv3-MobileNetV1结构的student网络进行蒸馏。
COCO数据集作为目标检测任务的训练目标难度更大,意味着teacher网络会预测出更多的背景bbox,如果直接用teacher的预测输出作为student学习的`soft label`会有严重的类别不均衡问题。解决这个问题需要引入新的方法,详细背景请参考论文:[Object detection at 200 Frames Per Second](https://arxiv.org/abs/1805.06361)。
为了确定蒸馏的对象,我们首先需要找到student和teacher网络得到的`x,y,w,h,cls,objectness`等Tensor,用teacher得到的结果指导student训练。具体实现可参考[代码](../../../ppdet/slim/distill_loss.py)
| 模型 | 方案 | 输入尺寸 | epochs | Box mAP | 配置文件 | 下载链接 |
| :---------------: | :---------: | :----: | :----: |:-----------: | :--------------: | :------------: |
| YOLOv3-ResNet34 | teacher | 608 | 270e | 36.2 | [config](../../yolov3/yolov3_r34_270e_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/yolov3_r34_270e_coco.pdparams) |
| YOLOv3-MobileNetV1 | student | 608 | 270e | 29.4 | [config](../../yolov3/yolov3_mobilenet_v1_270e_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_coco.pdparams) |
| YOLOv3-MobileNetV1 | distill | 608 | 270e | 31.0(+1.6) | [config](../../yolov3/yolov3_mobilenet_v1_270e_coco.yml),[slim_config](./yolov3_mobilenet_v1_coco_distill.yml) | [download](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_distill.pdparams) |
快速开始
```shell
# 单卡训练(不推荐)
python tools/train.py -c configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml --slim_config configs/slim/distill/yolov3_mobilenet_v1_coco_distill.yml
# 多卡训练
python -m paddle.distributed.launch --log_dir=logs/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml --slim_config configs/slim/distill/yolov3_mobilenet_v1_coco_distill.yml
# 评估
python tools/eval.py -c configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_distill.pdparams
# 预测
python tools/infer.py -c configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_distill.pdparams --infer_img=demo/000000014439_640x640.jpg
```
- `-c`: 指定模型配置文件,也是student配置文件。
- `--slim_config`: 指定压缩策略配置文件,也是teacher配置文件。
## FGD模型蒸馏
FGD全称为[Focal and Global Knowledge Distillation for Detectors](https://arxiv.org/abs/2111.11837v1),是目标检测任务的一种蒸馏方法,FGD蒸馏分为两个部分`Focal`和`Global`。`Focal`蒸馏分离图像的前景和背景,让学生模型分别关注教师模型的前景和背景部分特征的关键像素;`Global`蒸馏部分重建不同像素之间的关系并将其从教师转移到学生,以补偿`Focal`蒸馏中丢失的全局信息。试验结果表明,FGD蒸馏算法在基于anchor和anchor free的方法上能有效提升模型精度。
在PaddleDetection中,我们实现了FGD算法,并基于RetinaNet算法进行验证,实验结果如下:
| 模型 | 方案 | 输入尺寸 | epochs | Box mAP | 配置文件 | 下载链接 |
| ----------------- | ----------- | ------ | :----: | :-----------: | :--------------: | :------------: |
| RetinaNet-ResNet101| teacher | 1333x800 | 2x | 40.6 | [config](../../retinanet/retinanet_r101_fpn_2x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/retinanet_r101_fpn_2x_coco.pdparams) |
| RetinaNet-ResNet50 | student | 1333x800 | 2x | 39.1 | [config](../../retinanet/retinanet_r50_fpn_2x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/retinanet_r50_fpn_2x_coco.pdparams) |
| RetinaNet-ResNet50 | FGD | 1333x800 | 2x | 40.8(+1.7) | [config](../../retinanet/retinanet_r50_fpn_2x_coco.yml),[slim_config](./retinanet_resnet101_coco_distill.yml) | [download](https://paddledet.bj.bcebos.com/models/retinanet_r101_distill_r50_2x_coco.pdparams) |
快速开始
```shell
# 单卡训练(不推荐)
python tools/train.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml --slim_config configs/slim/distill/retinanet_resnet101_coco_distill.yml
# 多卡训练
python -m paddle.distributed.launch --log_dir=logs/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml --slim_config configs/slim/distill/retinanet_resnet101_coco_distill.yml
# 评估
python tools/eval.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/retinanet_r101_distill_r50_2x_coco.pdparams
# 预测
python tools/infer.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/retinanet_r101_distill_r50_2x_coco.pdparams --infer_img=demo/000000014439_640x640.jpg
```
- `-c`: 指定模型配置文件,也是student配置文件。
- `--slim_config`: 指定压缩策略配置文件,也是teacher配置文件。
## CWD模型蒸馏
CWD全称为[Channel-wise Knowledge Distillation for Dense Prediction*](https://arxiv.org/pdf/2011.13256.pdf),通过最小化教师网络与学生网络的通道概率图之间的 Kullback-Leibler (KL) 散度,使得在蒸馏过程更加关注每个通道的最显著的区域,进而提升文本检测与图像分割任务的精度。在PaddleDetection中,我们实现了CWD算法,并基于GFL和RetinaNet模型进行验证,实验结果如下:
| 模型 | 方案 | 输入尺寸 | epochs | Box mAP | 配置文件 | 下载链接 |
| ----------------- | ----------- | ------ | :----: | :-----------: | :--------------: | :------------: |
| RetinaNet-ResNet101| teacher | 1333x800 | 2x | 40.6 | [config](../../retinanet/retinanet_r101_fpn_2x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/retinanet_r101_fpn_2x_coco.pdparams) |
| RetinaNet-ResNet50 | student | 1333x800 | 2x | 39.1 | [config](../../retinanet/retinanet_r50_fpn_2x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/retinanet_r50_fpn_2x_coco.pdparams) |
| RetinaNet-ResNet50 | CWD | 1333x800 | 2x | 40.5(+1.4) | [config](../../retinanet/retinanet_r50_fpn_2x_coco.yml),[slim_config](./retinanet_resnet101_coco_distill_cwd.yml) | [download](https://paddledet.bj.bcebos.com/models/retinanet_r50_fpn_2x_coco_cwd.pdparams) |
| GFL_ResNet101-vd| teacher | 1333x800 | 2x | 46.8 | [config](../../gfl/gfl_r101vd_fpn_mstrain_2x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/gfl_r101vd_fpn_mstrain_2x_coco.pdparams) |
| GFL_ResNet50 | student | 1333x800 | 1x | 41.0 | [config](../../gfl/gfl_r50_fpn_1x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/gfl_r50_fpn_1x_coco.pdparams) |
| GFL_ResNet50 | CWD | 1333x800 | 2x | 44.0(+3.0) | [config](../../gfl/gfl_r50_fpn_1x_coco.yml),[slim_config](./gfl_r101vd_fpn_coco_distill_cwd.yml) | [download](https://bj.bcebos.com/v1/paddledet/models/gfl_r50_fpn_2x_coco_cwd.pdparams) |
快速开始
```shell
# 单卡训练(不推荐)
python tools/train.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml --slim_config configs/slim/distill/retinanet_resnet101_coco_distill_cwd.yml
# 多卡训练
python -m paddle.distributed.launch --log_dir=logs/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml --slim_config configs/slim/distill/retinanet_resnet101_coco_distill_cwd.yml
# 评估
python tools/eval.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/retinanet_r50_fpn_2x_coco_cwd.pdparams
# 预测
python tools/infer.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/retinanet_r50_fpn_2x_coco_cwd.pdparams --infer_img=demo/000000014439_640x640.jpg
# 单卡训练(不推荐)
python tools/train.py -c configs/gfl/gfl_r50_fpn_1x_coco.yml --slim_config configs/slim/distill/gfl_r101vd_fpn_coco_distill_cwd.yml
# 多卡训练
python -m paddle.distributed.launch --log_dir=logs/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/gfl/gfl_r50_fpn_1x_coco.yml --slim_config configs/slim/distill/gfl_r101vd_fpn_coco_distill_cwd.yml
# 评估
python tools/eval.py -c configs/gfl/gfl_r50_fpn_1x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/gfl_r50_fpn_2x_coco_cwd.pdparams
# 预测
python tools/infer.py -c configs/gfl/gfl_r50_fpn_1x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/gfl_r50_fpn_2x_coco_cwd.pdparams --infer_img=demo/000000014439_640x640.jpg
```
- `-c`: 指定模型配置文件,也是student配置文件。
- `--slim_config`: 指定压缩策略配置文件,也是teacher配置文件。
## LD模型蒸馏
LD全称为[Localization Distillation for Dense Object Detection](https://arxiv.org/abs/2102.12252),将回归框表示为概率分布,把分类任务的KD用在定位任务上,并且使用因地制宜、分而治之的策略,在不同的区域分别学习分类知识与定位知识。在PaddleDetection中,我们实现了LD算法,并基于GFL模型进行验证,实验结果如下:
| 模型 | 方案 | 输入尺寸 | epochs | Box mAP | 配置文件 | 下载链接 |
| ----------------- | ----------- | ------ | :----: | :-----------: | :--------------: | :------------: |
| GFL_ResNet101-vd| teacher | 1333x800 | 2x | 46.8 | [config](../../gfl/gfl_r101vd_fpn_mstrain_2x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/gfl_r101vd_fpn_mstrain_2x_coco.pdparams) |
| GFL_ResNet18-vd | student | 1333x800 | 1x | 36.6 | [config](../../gfl/gfl_r18vd_1x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/gfl_r18vd_1x_coco.pdparams) |
| GFL_ResNet18-vd | LD | 1333x800 | 1x | 38.2(+1.6) | [config](../../gfl/gfl_slim_ld_r18vd_1x_coco.yml),[slim_config](./gfl_ld_distill.yml) | [download](https://bj.bcebos.com/v1/paddledet/models/gfl_slim_ld_r18vd_1x_coco.pdparams) |
快速开始
```shell
# 单卡训练(不推荐)
python tools/train.py -c configs/gfl/gfl_slim_ld_r18vd_1x_coco.yml --slim_config configs/slim/distill/gfl_ld_distill.yml
# 多卡训练
python -m paddle.distributed.launch --log_dir=logs/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/gfl/gfl_slim_ld_r18vd_1x_coco.yml --slim_config configs/slim/distill/gfl_ld_distill.yml
# 评估
python tools/eval.py -c configs/gfl/gfl_slim_ld_r18vd_1x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/gfl_slim_ld_r18vd_1x_coco.pdparams
# 预测
python tools/infer.py -c configs/gfl/gfl_slim_ld_r18vd_1x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/gfl_slim_ld_r18vd_1x_coco.pdparams --infer_img=demo/000000014439_640x640.jpg
```
- `-c`: 指定模型配置文件,也是student配置文件。
- `--slim_config`: 指定压缩策略配置文件,也是teacher配置文件。
## PPYOLOE模型蒸馏
PaddleDetection提供了对PPYOLOE+ 进行模型蒸馏的方案,结合了logits蒸馏和feature蒸馏。
| 模型 | 方案 | 输入尺寸 | epochs | Box mAP | 配置文件 | 下载链接 |
| ----------------- | ----------- | ------ | :----: | :-----------: | :--------------: | :------------: |
| PP-YOLOE+_x | teacher | 640 | 80e | 54.7 | [config](../../ppyoloe/ppyoloe_plus_crn_x_80e_coco.yml) | [model](https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_x_80e_coco.pdparams) |
| PP-YOLOE+_l | student | 640 | 80e | 52.9 | [config](../../ppyoloe/ppyoloe_plus_crn_l_80e_coco.yml) | [model](https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_l_80e_coco.pdparams) |
| PP-YOLOE+_l | distill | 640 | 80e | **54.0(+1.1)** | [config](../../ppyoloe/distill/ppyoloe_plus_crn_l_80e_coco_distill.yml),[slim_config](./ppyoloe_plus_distill_x_distill_l.yml) | [model](https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_l_80e_coco_distill.pdparams) |
| PP-YOLOE+_l | teacher | 640 | 80e | 52.9 | [config](../../ppyoloe/ppyoloe_plus_crn_l_80e_coco.yml) | [model](https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_l_80e_coco.pdparams) |
| PP-YOLOE+_m | student | 640 | 80e | 49.8 | [config](../../ppyoloe/ppyoloe_plus_crn_m_80e_coco.yml) | [model](https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_m_80e_coco.pdparams) |
| PP-YOLOE+_m | distill | 640 | 80e | **51.0(+1.2)** | [config](../../ppyoloe/distill/ppyoloe_plus_crn_m_80e_coco_distill.yml),[slim_config](./ppyoloe_plus_distill_l_distill_m.yml) | [model](https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_m_80e_coco_distill.pdparams) |
快速开始
```shell
# 单卡训练(不推荐)
python tools/train.py -c configs/ppyoloe/distill/ppyoloe_plus_crn_l_80e_coco_distill.yml --slim_config configs/slim/distill/ppyoloe_plus_distill_x_distill_l.yml
# 多卡训练
python -m paddle.distributed.launch --log_dir=logs/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/ppyoloe/distill/ppyoloe_plus_crn_l_80e_coco_distill.yml --slim_config configs/slim/distill/ppyoloe_plus_distill_x_distill_l.yml
# 评估
python tools/eval.py -c configs/ppyoloe/distill/ppyoloe_plus_crn_l_80e_coco_distill.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_l_80e_coco_distill.pdparams
# 预测
python tools/infer.py -c configs/ppyoloe/distill/ppyoloe_plus_crn_l_80e_coco_distill.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_l_80e_coco_distill.pdparams --infer_img=demo/000000014439_640x640.jpg
```
- `-c`: 指定模型配置文件,也是student配置文件。
- `--slim_config`: 指定压缩策略配置文件,也是teacher配置文件。
## 引用
```
@article{mehta2018object,
title={Object detection at 200 Frames Per Second},
author={Rakesh Mehta and Cemalettin Ozturk},
year={2018},
eprint={1805.06361},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
@inproceedings{yang2022focal,
title={Focal and global knowledge distillation for detectors},
author={Yang, Zhendong and Li, Zhe and Jiang, Xiaohu and Gong, Yuan and Yuan, Zehuan and Zhao, Danpei and Yuan, Chun},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={4643--4652},
year={2022}
}
@Inproceedings{zheng2022LD,
title={Localization Distillation for Dense Object Detection},
author= {Zheng, Zhaohui and Ye, Rongguang and Wang, Ping and Ren, Dongwei and Zuo, Wangmeng and Hou, Qibin and Cheng, Mingming},
booktitle={CVPR},
year={2022}
}
@inproceedings{shu2021channel,
title={Channel-wise knowledge distillation for dense prediction},
author={Shu, Changyong and Liu, Yifan and Gao, Jianfei and Yan, Zheng and Shen, Chunhua},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
pages={5311--5320},
year={2021}
}
```